12

I've listened to many interviews and lectures of Alain Connes, in which he says something which goes roughly as follows

"Every non-commutative algebra has its own time (evolution of), by which I mean a one-parameter group."

I find this statement somewhat mysterious and intriguing at the same time.

Question. What is the precise statement of this result and how to can this proven explicitly even for a simple scenario like, say, the von Neumann algebra $M_2(\mathbb C)$ of $2 \times 2$ complex matrices ?

Disclaimer. I have essentially no knowledge of non-comutative geometry, etc. I'd appreciate a very simple construction / proof; nothing too fancy. Thanks.

Dmitri Pavlov
  • 36,141
dohmatob
  • 6,726
  • 2
    I have seen this quote before, with varying levels of precision from Connes, and it is best thought of as being in the tradition of Francophone epigrams, or "poetically true" rather than "literally true" – Yemon Choi Mar 14 '21 at 19:40
  • 7
    In particular, the given claim applied to $M_2({\bf C})$ just yields a time evolution that is trivial, i.e. staying the same for all time. What Connes is referring to is, AFAICT from various readings, the so-called "modular flow" of a von Neumann algebra. The technical details are outlined at https://en.wikipedia.org/wiki/Tomita%E2%80%93Takesaki_theory#The_Connes_cocycle - I'm aware that this doesn't meet your request of a simple account but it might provide some of the right words to search for or ask about. – Yemon Choi Mar 14 '21 at 19:45
  • @YemonChoi Thanks for the valuable input. Do you have soft reference on "flows on von Neumann algebras" ? Also, Do you see a mild modification of my matrix example for which the flow is not trivial (i.e non-static) ? Thanks in advance. – dohmatob Mar 14 '21 at 21:37
  • @YemonChoi: It is possible to give a competely precise (i.e., literally true) expression of the modular flow as an actual action of a group, see my answer. – Dmitri Pavlov Mar 14 '21 at 22:53
  • 1
    @DmitriPavlov I was aware that this can be done, although not of the details that you provide in your impressive answer; but I stand by my first comment in light of the fact that not every noncommutative algebra over C is a von Neumann algebra, as Connes is clearly well-aware. – Yemon Choi Mar 15 '21 at 00:34
  • 1
    @dohmatob All von Neumann algebras with a faithful normal tracial state will have trivial modular flow. In particular this is true for all finite-dimensional von Neumann algebras (= finite sums of finite matrix algebras). Hence why I find the slogan of Connes that you have quoted somewhat exasperating (and this is even before we get into non-selfadjoint operator algebras, or Banach star-algebras that are not Cstar algebras, etc). As Dmitri's answer indicates at the end, you need to go to the setting of Type III von Neumann algebras to get something nontrivial. – Yemon Choi Mar 15 '21 at 00:36
  • @YemonChoi I see. Thanks! – dohmatob Mar 15 '21 at 00:57
  • See my answer posted 6 years ago: https://mathoverflow.net/a/186527/34538 – Sebastien Palcoux Mar 16 '21 at 21:31
  • @SebastienPalcoux I cannot comment on the vN aspects of your answer but your attempt to get a vN algebra from an associative ${\mathbb C}$-algebra seems to have lots of gaps in it. What do you do if your algebra does not have a countable Hamel basis? Why do you think that your construction is independent of the choice of basis? etc. etc. – Yemon Choi Mar 19 '21 at 15:03
  • 1
    Sometimes I think that people whose only exposure to the NC world is through a vN background or by reading the sweeping statements from NCG, could do with reading and working through Howie's Fundamentals of Semigroup Theory, or reading the classical literature on nest algebras – Yemon Choi Mar 19 '21 at 15:06
  • @YemonChoi: your last two comments would be more suitable below the answer I mentioned. By the way, you are right in the sense that I should not write « the » vN algebra $\mathcal{M}$ generated by $\mathcal{A}$ but « a » vN algebra... In addition, it is possible that, in some cases, $\mathcal{A}$ does not embed into $\mathcal{M}$ (for example if the multiplication operators of some basis elements are unbounded). – Sebastien Palcoux Mar 19 '21 at 17:16

1 Answers1

13

Given any von Neumann algebra $M$, we can define its noncommutative $\def\L{{\cal L}} \L^p$-spaces $\L^p(M)$ for any $\def\C{{\bf C}} p∈\C$ such that $\Re p≥0$. Here I use the notation $\L^p:={\rm L}^{1/p}$, where the right side is the usual notation from measure theory and real analysis. (This notation is explained in more detail in another answer.)

We have $\L^0(M)≅M$, $\L^1(M)≅M_*$, and $\L^{1/2}(M)$ is the standard form of $M$ due to Haagerup. The spaces $\L^p(M)$ for all $p∈\C$ form a $\C$-graded *-algebra, where the involution is a $\C$-antilinear map $\L^p(M)→\L^{\bar p}(M)$. The multiplication is well-defined because of Hölder's inequality, which in this context says that we have a map $$\L^p(M)⊗_\C\L^q(M)→\L^{p+q}(M).$$

In fact, we can do better: the induced map $$\L^p(M)⊗_{\L^0(M)}\L^q(M)→\L^{p+q}(M)$$ is an isomorphism for any $p,q∈\C$ such that $\Re p≥0$, $\Re q≥0$. Here the tensor product on the left side is purely algebraic, it happens to be automatically complete.

In particular, the bimodules $\L^p(M)$ are invertible for any $\def\I{{\bf I}} p∈\I=\{z∈\C\mid \Re z=0\}$, with the inverse being the bimodule $\L^{-p}(M)=\L^{\bar p}(M)$.

Thus, we have a morphism of 2-groups $$\def\VNA{{\sf VNA}} \L(M)\colon\I→\VNA^⨯_M,$$ where $\VNA$ denotes the bicategory of von Neumann algebras, W*-bimodules, and (bounded) intertwiners, $\VNA^⨯$ denotes its maximal 2-subgroupoid, i.e., we only take invertible bimodules (alias Morita equivalences) and invertible intertwiners, and $\VNA^⨯_M$ denotes the 2-subgroupoid of $\VNA^⨯$ on a single object, namely, $M$, whose endomorphisms form a monoidal groupoid, i.e., a 2-group.

The bicategory $\VNA^⨯$ is analogous to the bicategory of Lie groupoids, Morita equivalences (given by invertible bibundles), and invertible morphisms of bibundles. In particular, 1-morphisms in $\VNA^⨯$ can be seen as invertible maps of corresponding noncommutative measurable spaces in the same way as Morita equivalences of Lie groupoids can be seen as invertible maps of stacks corresponding to these Lie groupoids.

Thus, the canonical homomorphism of 2-groups $\L(M)\colon\I→\VNA^⨯_M$ equips every von Neumann algebra $M$ with a canonical action of the Lie group $\I=i{\bf R}$ of imaginary numbers. This is precisely the Tomita–Takesaki modular flow, expressed in a canonical manner without arbitrary choices.

If $M$ is a type III von Neumann algebra, this action is not isomorphic to the trivial action.

This can be most easily seen as follows. The trivial action sends all $t∈\I$ to the $M$-$M$-bimodule $M$, with the obvious trivial coherence data. If α is an isomorphism from the trivial action to the action $\L(M)$, then for each $t∈\I$ we have an isomorphism of $M$-$M$-bimodules $$α_t\colon M→\L^t(M).$$ Such an isomorphism is uniquely characterized by the element $α_t(1)∈\L^t(M)$, which must be a central element of support 1. Now the relations imposed by the definition of a homomorphism of 2-groups tell us that $$α_s(1)α_t(1)=α_{s+t}(1)∈\L^{s+t}(M)$$ and $α_0(1)=1∈\L^0(M)=M$. Thus, $$t↦α_t(M)∈\L^t(M)$$ is a one-parameter group of unitary elements of $\L(M)$. Such one-parameter groups are in bijection with faithful semifinite normal weights $μ$ on $M$: given $μ$, we set $$α_t(1)=μ^t∈\L^t(M).$$ The elements $α_t(1)$ are central in $\L^t(M)$ if and only if $μ$ is a trace. (If $μ$ is finite, then $μ(ab-ba)=0$ can be rewritten as $(μa-aμ)(b)=0$, i.e., $aμ=μa$ for all $a∈M$. That is to say, $α_1(1)=μ^1=μ∈\L^1(M)$ is a central element.)

Thus, the action $\L(M)$ is isomorphic to the trivial action if and only if $M$ admits a faithful semifinite normal trace, i.e., $M$ is a semifinite von Neumann algebra, equivalently, a direct integral of type I and type II factors.

Hence, if $M$ is a type III algebra, the action is nontrivial.

This approach also makes it clear the extent to which the traditional modular automorphism groups depend on the choice of a faithful semifinite normal weight. Given $t\in\I$ and a faithful semifinite normal weight $μ$, the traditional Tomita–Takesaki modular automorphism is $$σ_μ^t\colon M→M,\qquad x↦σ_μ^t(x)=μ^t x μ^{-t}.$$

We have a canonical homomorphism of 2-groups $$\def\Aut{\mathop{\rm Aut}} T\colon \Aut(M) → \VNA^⨯_M$$ that sends $σ∈\Aut(M)$ to the $M$-$M$-bimodule $T(σ)$. This is indeed a homomorphism because we have canonical isomorphisms $T(σ)⊗_M T(σ')≅T(σσ')$ that satisfy the relevant coherence conditions.

Now consider the $M$-$M$-bimodule $M_{μ,t}=T(σ_μ^t)$, where the right action is twisted by $σ_μ^t$: $a⋅m⋅b=amσ_μ^t(b)=amμ^t b μ^{-t}$. We have a canonical isomorphism of $M$-$M$-bimodules $$ρ_{μ,t}\colon M_{μ,t}→\L^t(M), \qquad m↦mμ^t.$$

Furthermore, the isomorphisms $ρ_{μ,t}$ are compatible with tensor products: $$ρ_{μ,s}⊗_M ρ_{μ,t}≅ρ_{μ,s+t}.$$

This amounts to saying that for any faithful semifinite normal weight $μ$, the composition of the homomorphism of groups $$σ_μ\colon \I → \Aut(M)$$ with the homomorphism of 2-groups $$T\colon\Aut(M) → \VNA^⨯_M$$ that twists the right action is canonically isomorphic (via the isomorphism $ρ_μ$) to the homomorphism of 2-groups $$\L(M)\colon \I → \VNA^⨯_M.$$

Observe that both $σ_μ$ and the isomorphism $$ρ_μ\colon T∘σ_μ → \L(M)$$ depend on the choice of a faithful semifinite normal weight $μ$, but the homomorphism $\L(M)$ does not. Thus, $\L(M)$ expresses the modular automorphism group in a coordinate-free way.

Dmitri Pavlov
  • 36,141
  • Awesome! Is there some book to read more about this stuff? Specifically I would like to know whether there are some other invertible bimodules and where to find them, if yes. Also, many thanks for lots of fantastic material on your home page. – მამუკა ჯიბლაძე Mar 15 '21 at 08:15
  • @მამუკაჯიბლაძე: Invertible bimodules (i.e., Morita equivalences) of von Neumann algebras have been studied a lot. Any such bimodule can be obtained as a composition of an amplification (an M-M⊗B(H)-bimodule M⊗H, for a Hilbert space H) and a reduction (an M-pMp-bimodule Mp, for a projection p∈M with central support 1). Morita equivalent von Neumann algebras have isomorphic centers, and Morita equivalences respect direct integral decompositions of von Neumann algebras, so it suffices to study Morita equivalences of factors. – Dmitri Pavlov Mar 15 '21 at 16:14
  • @მამუკაჯიბლაძე: All type I factors are Morita equivalent. If Morita equivalent type III factors act on a separable Hilbert space, they are isomorphic. Any type II_∞ factor is Morita equivalent to a type II_1 factor. Finally, for a type II_1 factor M any Morita equivalent type II_1 factor has the form End(P), where P is a finitely generated module over M, characterized by its dimension dim(P)∈(0,∞); we have P=M^n⊕pMp, where n is the integer part of dim(P) and p is a projection in M of the trace dim(P)−n. – Dmitri Pavlov Mar 15 '21 at 16:17
  • 2
    @მამუკაჯიბლაძე: The canonical reference is Marc Rieffel's paper Morita equivalence for C-algebras and W-algebras. – Dmitri Pavlov Mar 15 '21 at 16:19
  • Thank you. But Rieffel's paper hardly contains anything about Picard group(oid)s, does it? So can one say that the groupoid of Morita self-equivalences of a vN-algebra $M$ is some kind of categorified real vector space, equipped with a distinguished 1-dimensional subspace $\mathscr L^{i\mathbb R}(M)$? – მამუკა ჯიბლაძე Mar 15 '21 at 17:23
  • 1
    @მამუკაჯიბლაძე: I would say that the Picard 2-group of a von Neumann algebra categorifies the unitary group of a Hilbert space. Unlike the unitary group, it is equipped with a canonical 1-parameter subgroup. I don't think there is much literature on Picard 2-groups of von Neumann algebras, since operator algebraists rarely talk about categories. However, André Henriques's papers probably say something about Picard 2-groups. André is actually here on MathOverflow, we could try to ask him. – Dmitri Pavlov Mar 15 '21 at 18:11
  • Very nice! I'm confused, though -- if I'm following correctly, what we have here is a 1-parameter subgroup $\mathcal L(M)$ of the Picard group $Pic^0(vNA_M) = vNA^\times_M$, carrying $t \mapsto \mathcal L_{it}(M)$, but isn't the modular flow rather a 1-parameter subgroup of $Pic^1(vNA_M) = Aut(M)$? Perhaps I'm missing something basic related to the fact that the modular flow is only well-defined up to inner automorphism? (By the way, I'm surprised to see so much Latex in your post!) – Tim Campion Jul 26 '21 at 17:16
  • @TimCampion: I added a few paragraphs about this at the end, it really is the same thing. I actually did not use any LaTeX in my post, just plain TeX. Concerning TeX, I predicted that hordes of poorly qualified self-appointed TeXifiers would descend on this post since it uses lots of formulas, so I preempted their attack by using TeX myself. But even today, the Chrome/Chromium browser does not support formulas natively (only through JavaScript), so the problem is still not solved. – Dmitri Pavlov Jul 27 '21 at 00:04